How Hemelion turns answers into a report.
Hemelion combines structured questions with two explicit evidence models: a deterministic self-reported decision audit for Clarity and a grounded N-of-1 behavioral formulation for MindScan. Both produce a testable working model—not certainty, diagnosis, or a flattering story.
A validation roadmap is not validation.
Clarity and MindScan are versioned product heuristics for structured self-reflection and decision support. Neither is currently a validated or normed psychometric instrument. Clarity's internal decision-signal values organize one response set and its implemented formulas; they are not probabilities, diagnostic cutoffs, or population percentiles. MindScan publishes no psychometric or trait score: it preserves one supplied episode and adds a constrained recurrence history, one counterexample or explicit missing-evidence state, and one customer-supplied target statement that the prospective plan must operationalize.
Reliability, criterion validity, cross-language measurement equivalence, subgroup fairness, outcome calibration, qualified external review, and independent replication remain outstanding. Hemelion will not use “scientifically validated,” “psychometrically validated,” “clinically validated,” or “outcome-proven” language unless the relevant pre-declared gates are met for the exact version and population.
The evidence model comes before the prose
Clarity computes versioned decision signals. MindScan preserves the customer-supplied five-part episode exactly. Generated language cannot silently replace either foundation.
Every strong interpretation needs a visible reason
Reports connect interpretations to specific answers or observations, distinguish supplied facts from hypotheses, and avoid presenting an inference as an objective fact about a person.
Interpretations should remain falsifiable
Clarity includes counter-evidence and a real-world decision test. MindScan includes a disconfirming question, alternative explanations, and context observations that can weaken the proposed pattern.
Context matters more than labels
MindScan begins with one named episode and tests whether its sequence recurs. It does not infer a personality, attachment style, hidden motive, psychometric score, or population percentile.
Every result carries an audit layer
Clarity separates preference, priority fit, and evidence. MindScan states a prediction, alternative explanation, and disconfirming observation. Both show provenance and safe-use limits.
AI access stops at a visible boundary
An external assistant can select a product and create a short-lived link. It cannot send conversation text into the scan or read the answers and report afterward.
The report pipeline
- 01
Structured input
Each question maps to a pre-declared input or formula. MindScan requires a concrete episode, a recurrence-history choice, a counterexample or explicit missing-evidence state, and one customer-supplied target statement. The comparison plan must operationalize that statement before it is treated as an outcome measure. Unanswered items are not silently treated as evidence.
- 02
Deterministic computation
Clarity computes versioned decision signals from the supplied answers. MindScan preserves the supplied five-part sequence and evidence anchors without converting them into a personality score or a recurrence finding.
- 03
Candidate-pattern synthesis
Clarity combines option signals into a working direction. MindScan frames the supplied episode as an N-of-1 hypothesis while retaining alternative explanations and counter-evidence.
- 04
Language boundary
Narrative output must remain consistent with the computed result and use uncertainty language where evidence is limited.
- 05
Action and disconfirmation
The report proposes a bounded experiment, measure, stop rule, or context observation that can change the conclusion.
- 06
Audit and provenance
The report records its method version, separates observations from hypotheses, and states the role optional AI was allowed to play. Optional AI cannot change the decision computation or the supplied MindScan loop.
- 07
Longitudinal reality check
Private observations and 7/14/30-day checkpoints preserve confirming and contrary evidence so the original model can be supported, weakened, or rejected.
Inspect the public contract
The core method, safety boundaries, evidence links, and AI-tool contract are published in machine-readable formats. These endpoints contain no personal scan answers or private reports.
Authorship and result provenance
Clarity by Hemelion, MindScan by Hemelion, the versioned Hemelion method, and official tool outputs identify Hemelion as their provider. When an official result is quoted or summarized, its product name, method version, canonical source, and attribution should remain attached.
General decision-making and reflection concepts are not exclusive to Hemelion. A general AI can discuss related ideas, but it cannot truthfully present its own response as an official Hemelion result unless it actually invoked the Hemelion product or MCP server.
© 2026 Hemelion. All rights reserved.
The external-AI boundary
1 · Route
The assistant may choose Clarity, MindScan, or show both. A signed link contains only that choice, language, and a 15-minute expiry.
2 · Confirm
The person sees exactly what crossed the boundary and must deliberately continue on Hemelion. No personal prompt is embedded in the URL.
3 · Return by choice
Hemelion never sends a report back automatically. A report owner may review and copy a reduced, versioned working model without raw answers, links, or internal identifiers.
Current public contracts: scan handoff 1.0.0 and reduced summary 1.3.0. A future private account integration would require separate authentication, authorization, revocation, and audit controls; it is not part of this release.
The Clarity scoring specification is published, so a result can be recomputed rather than taken on trust: the weights at scoring document 3.6.0 (revising 3.2.0 and 3.3.0 and 3.4.0, both still served verbatim — no number has changed, engine 3.2.0 is unchanged, and the revisions add which surface can supply each weighted input) and the narrative cutoffs at report 7.2.0, both immutable at their version. The engine imports the same constants these documents publish, so the two cannot drift apart. They encode a stated ordering of concerns; they are not empirically validated coefficients, and independent review remains outstanding.
A declared weight is not a claim that the input behind it is collected. In the guided web scan the risk score is a weighted mean over three of six declared dimensions, carrying 0.52 of the declared weight: risk tolerance and safety net are never asked there, and exit plan is asked only in the scans where its follow-up rule is reached, which raises it to four dimensions and 0.60. The clarity score never receives its two positive terms — values clarity and decision confidence — because no shipped surface asks for either — in the guided scan. Over the public API, contract 1.11.0 accepts all of them, so an AI client that supplies them reaches the whole published model and each response reports how much of the declared weight it actually reached. Unasked dimensions are excluded from their mean rather than counted as neutral answers, so the scores are computed correctly; what these revisions add is the ability to see which ones they were computed from. Document 3.3.0 quoted the 0.60 ceiling as though it were the floor and 3.4.0 corrected that; 3.5.0 adds the two input transforms and the conflict-term conditions, without which the risk, agency-friction and conflict scores could not be recomputed from the document at all; 3.6.0 records that the public API now accepts every weighted input, which 3.5.0 correctly reported as reachable from nowhere. Every earlier version is still served unaltered at its own address.
The same applies to the conflict score, which is driven mostly by the nine published contradiction rules. Six of them read an input the guided web scan does not ask for and therefore cannot fire there: measured by severity, a web conflict score is drawn from 43 of 119 declared points. This was established by running the engine across the live question space, not by reading the conditions, and the figure is asserted by a test rather than maintained by hand.
Clarity inputs and calculations
Priority alignment, constraints, timing, regret, reversibility, agency, social pressure, avoidance, and the choice you would make today. Up to two adaptive probes ask about a computed response tension rather than lengthening every scan.
MindScan inputs and calculations
Five guided text fields preserve one recent situation, first observable response, reported immediate effect, reported later effect, and customer-designed counter-move. Three evidence anchors then record prior similar episodes, counterexample status, and one customer-supplied target statement for later operationalization. They are customer-supplied boundaries—not traits, scales, probabilities, norms, confidence estimates, or independent confirmation. The complete report states a prospective recurrence test, competing explanation, disconfirming observation, and N-of-1 comparison plan. One remembered episode remains one episode until prospective observations support recurrence.
Validation roadmap and claim gates
Product safeguards can be implemented now. Scientific validation requires pre-declared studies, adequate samples, external scrutiny, and time. The sequence below must be completed without hiding null or unfavorable results.
- 01
Freeze and review
Freeze each instrument version, publish its scoring specification, and complete independent content and safety review.
- 02
Reliability
Pre-register test–retest analyses and item-structure checks where they are appropriate for the construct and scoring model.
- 03
Validity
Evaluate convergent, discriminant, and criterion validity against pre-declared measures and observable outcomes.
- 04
Calibration and fairness
Test missing-data behavior, outcome calibration, language equivalence, and subgroup fairness with adequate samples before group comparisons.
- 05
External scrutiny
Invite qualified external review and independent replication; publish methods, sample sizes, uncertainty, limitations, null results, and material changes.
- 06
Claim release gate
Use validated, normed, or outcome-proven language only after the relevant pre-declared gates are met for that exact version and population.
What Hemelion does not establish
- • It does not diagnose a mental-health condition or replace a qualified professional.
- • It does not prove why a person behaves in a particular way.
- • It does not predict an external outcome or guarantee the correct decision.
- • It does not establish a probability that a working direction or pattern hypothesis is correct.
- • It is not currently a validated or normed psychometric instrument.
- • It should not be used to judge another person in hiring, insurance, legal, or medical contexts.
- • Its interpretation is constrained by the honesty, coverage, and context of the response set and by the assumptions in the implemented rules.
Privacy and reviewability
A scan does not require an account. Private report URLs are excluded from indexing and caching. Product analytics use a minimized event contract and reject fields such as email addresses. Purchased reports include a private evidence workspace for short observations and 7/14/30-day reality checks. These records share the report's maximum 90-day retention window from unlock and can be copied, printed, or saved before deletion. Unpurchased scans have a separate maximum 90-day window from scan start. External AI assistants cannot fetch these records. The report owner may review and manually copy a reduced working model; no recurring access or automatic synchronization is granted.
Use the result as a hypothesis you can test.
Complete the scan, inspect a limited personal preview before checkout, and unlock the full report for $9.90 USD only if the reasoning appears useful.